The evolution of the photofit

Not the usual suspects

IF YOU have ever tried to describe someone's face in detail, you will understand why the “photofit” composite images used by the police look so dodgy. In recent years, computerised “E-Fit” systems have helped improve the accuracy of these images by allowing witnesses to choose from a wider range of facial features. But researchers at the University of Stirling in Scotland found that, despite these improvements, people still have a hard time constructing recognisable faces—especially if there is a delay of more than a few hours between seeing the face and generating the composite.

The problem is that people tend to recognise faces as whole entities, rather than as separate features. So picking from a range of eyes, noses and mouths is not necessarily the most suitable approach, says Charlie Frowd, a psychologist at the University of Stirling. Next year, however, British police are expected to begin trials of a new computerised system, called EvoFIT, developed by Dr Frowd and his colleague Peter Hancock. It uses an evolutionary approach, known as a genetic algorithm, to “evolve” faces rather than piece them together.

The process is entirely non-verbal and takes a fraction of the normal time, claims Dr Frowd. A witness is shown an array of 60 different faces with random features. Having studied them closely, the witness is asked to choose the six images that most closely resemble the person they are trying to describe. These six are then used to generate another set of 60, by switching features between some of the images and by making random changes to others. The witness is then asked to repeat the task, whereupon a new batch of faces is generated, and so on. In evolutionary terms, this process is known as “genetic crossover and mutation”, and is a powerful way to search a large number of possibilities for a particular solution. By allowing the user to steer the selection process, the program is able to generate a good likeness for the original face after just a few cycles.

In one of its early versions, EvoFIT was used by police in Northamptonshire who were trying to catch a violent attacker. The attacker was never caught, but the senior investigating officer, Superintendent Paul Spick, says the witness involved found the new software much quicker and easier to use than traditional E-Fit systems. It was also more accurate: the final image caused the witness visible distress when displayed. A further important advantage of EvoFIT over traditional composites is that an image can be generated even if the witness can only provide a sketchy verbal description.

The researchers have since made a number of further improvements to their system and are now in the process of commercialising it in partnership with ABM, a British firm which supplies police forces with PROfit, one of the leading computerised composite systems. The images are more realistic, and the system can generate three-quarter angle views of the face, which are easier to distinguish than direct frontal views, says Dr Frowd.

His team has also found that by combining the images generated by different witnesses, or even from a single witness, it is possible to get an even better likeness. This could be particularly useful when multiple witnesses come forward and the police are unsure which of the images they produce is the most accurate. Another improvement has come from changing the image-selection process. In the latest version, witnesses are given a selection of face shapes to choose from before facial features are added. This makes it less likely that the correct facial features will be rejected simply because they are on the wrong-shaped face.

These modifications appear to make all the difference. In recent experiments, Dr Frowd and Dr Hancock compared EvoFIT with the computerised systems currently in use by asking volunteers to construct an image of a face two days after seeing it, and then showing the resulting image to people who were familiar with the person depicted. With traditional composites, the person was correctly identified about 4% of the time; the figure rose to 25% with EvoFIT images.

Next year ABM will perform in-house evaluations of the system and will then team up with police forces for further trials. Although things look promising so far, two things in particular need further testing, says Leslie Bowie, ABM's director of research. Do the EvoFIT images distract witnesses or contaminate their memories of the face they are trying to recall? And while EvoFIT has performed well in the calm conditions of the laboratory, how will it cope with real-life witnesses, who are often traumatised by their experiences? If it can address these two questions, EvoFIT could be the next step in the evolution of the photofit.